source stringclasses 1
value | split stringclasses 1
value | seeds listlengths 2 2 | scenes int64 100 300 | stages int64 4 4 | horizon int64 1.8k 1.8k | goal stringclasses 1
value | reached dict | every int64 80 80 ⌀ | subgoals_per_scene dict |
|---|---|---|---|---|---|---|---|---|---|
scripted nominal + recovery | val | [
100035,
101075
] | 100 | 4 | 1,795 | first step with k placed cubes | {
"stage1": 1,
"stage2": 1,
"stage3": 1,
"stage4": 1
} | 80 | {
"mean": 18.85,
"min": 18,
"max": 22
} |
scripted nominal + recovery | val | [
100035,
101075
] | 100 | 4 | 1,795 | first step with k placed cubes | {
"stage1": 1,
"stage2": 1,
"stage3": 1,
"stage4": 1
} | null | null |
scripted nominal + recovery | val | [
100035,
103240
] | 300 | 4 | 1,795 | first step with k placed cubes | {
"stage1": 1,
"stage2": 1,
"stage3": 1,
"stage4": 0.9766666666666667
} | 80 | {
"mean": 18.886666666666667,
"min": 18,
"max": 22
} |
CIS 6280 Stacking Datasets
Demonstration episodes, pre-encoded latents, on-policy candidate rollouts, and subgoal images for the CIS 6280 four-box stacking assignment. Pretrained world models and policies are in YongYong/cis6280-stack-world-models-v2.
What to Download
If you only need Stage 3 (policy training), Stage 4 (predictor finetuning), or Stage 5 (evaluation), you can skip the 40 raw episode shards (190 GB) and download only the derived files and subgoals (13 GB):
hf download YongYong/cis6280-stack-datasets-v2 --repo-type dataset --local-dir dataset \
--include "derived/*" "onpolicy-smwm-c3/*" "goals/*"
Contents
| Path | Description |
|---|---|
bc-noise-4cube-20k/ |
20,000 demonstration episodes in 40 tar shards (500 episodes each), recorded at 10 Hz with wrist-camera RGB rendered every 5 control steps (2 Hz). Collected on scenes 400000–419999 (15,990 train, 1,991 val, 2,019 test) with the privileged BC teacher bc/step-050000.pt; 10% of episodes use $\sigma = 0$ and 90% sample action noise $\sigma \in [0.005, 0.2]$ log-uniformly. |
derived/jepa-cache-4cube-20k/ |
Episode index (index.json) and packed 5-step action blocks (actions.f32, shape (7345132, 125), float32) across all 20,000 episodes. |
derived/latents-smwm-c3/ |
Every cached frame already encoded into a 192-D latent by our world model (latents.f16 and latents.json), so you can train a policy in Stage 3 without the raw images. |
onpolicy-smwm-c3/ |
On-policy candidate rollouts of policy/df-smwm-c1-sigma04/policy.pt used in Stage 4 predictor finetuning (33,984 training states and 6,208 validation states, each with 16 candidate 16-block rollouts). |
goals/ |
Subgoal images taken every 80 control steps (8 s) from a scripted reference run on the first 300 validation-split scenes at or after seed 100000 (100035–103240). scripted-val300-every80/goals.npz covers all 300 benchmark scenes; scripted-val100-every80/goals.npz covers the first 100 scenes for Google Colab. |
Rebuilding the Raw Image File for Stage 2
The raw image memmap (frames.u8, about 722 GB) is not on Hugging Face. You only need to build it if you are pretraining a new world-model encoder in Stage 2:
cd dataset/bc-noise-4cube-20k && for f in episodes-*.tar; do tar xf "$f"; done && cd ../..
python -m assignment.training.data.cache --data dataset/bc-noise-4cube-20k --out cache-4cube-20k
- Downloads last month
- 221